FAQ
Which criterion for optimality should I choose?
Original question from a Chemical Engineer:
“I am setting up a response surface method (RSM) process optimization to establish the boundaries of my system and find the peak of performance. I usually go with your default of I-optimality for custom designs. However, this optimality focuses more on the extremes than modified distance or distance. To fit my purpose, what optimality should I use?”
Answer :
I do not completely agree that I-optimality tends to be too extreme. It actually does a lot better at putting points in the interior than D-optimality as shown in Figure 2 at Practical Aspects for Designing Statistically Optimal Experiments. For that reason, Stat-Ease software defaults to I-optimal design for optimization and D-optimal for screening (process factorials or extreme-vertices mixture).
Also, keep in mind that, if you go with I-optimality and keep the 5 lack-of-fit points added by default using a distance-based algorithm, you will get an outstanding combination of ideally located model points plus other points that fill in the gaps.
However, if you prefer a design that spaces out all your points, go with the modified distance design, which maintains the ability to fit the model you specify (defaulting to quadratic). Purely distance-based designs are best for ‘black-box’ computer simulations (not your situation being a hands-on experimenter). In that case, consider upgrading to Stat-Ease® 360 software, which offers a broader choice of space-filling designs.
PS: I am working on a blog post that graphically illustrates how choices on our optimal custom design builder affect point location. Stay tuned to the Stat-Ease blog page for this post.
(Learn more about optimal design by enrolling in the next Modern DOE for Process Optimization public workshop.)
